Camera-and-microphone glasses record the among the richest first-person evidence streams any consumer device produces: what the wearer sees, where, and with which objects, documents, and people present. Today that stream is evaluated by capture quality and recall — transcription accuracy, scene search, question answering over footage. What no current evaluation tests is whether egocentric evidence makes a system measurably better at forecasting what this wearer does next.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
Personal AI products are converging along one ladder: recording → structured memory → longitudinal evidence → target-specific predictive capability → validated assistance, with the predictive-capability rung measured by TargetSpace. Smart glasses entered as hands-free meeting intelligence — transcribe what was said, caption what was seen — and current flagship features are memory-shaped: “where did I leave my keys”, “who was that person”, “summarize my morning”. That is retrieval over an egocentric archive. Memory is not understanding; a system that surfaces what the wearer saw is not yet a system that forecasts what the wearer will do.
Current rung: structured memory. Egocentric capture plus retrieval and summarization over it. Next rung: longitudinal evidence — treating the wearer's visual history as ordered evidence about their state transitions, the input to the sealed forecasts TargetSpace scores. Whether that history buys forecasting skill is an open, testable question, not a premise.
In the TargetSpace evidence-tier scheme, smart glasses contribute L3–L4 evidence — screen and egocentric audiovisual context: what is seen, where, and which objects, documents, and people are present — layered on top of L2 audio from the on-device microphones. Tasks come from the flagship TS-Personal track: sealed forecasts about a specific wearer's next observable state transition, resolved deterministically after the fact. The marginal value of each tier is established by evidence-tier ablation: run the same sealed tasks with vision on and off, and report what the egocentric stream actually bought. Passive visual capture outperforming lighter channels is a hypothesis this protocol tests, not an assumption it makes.
At a sealed time t inside a defined work session, forecast the wearer's next transition using only egocentric context observed up to t: continue the current task, switch to a different task, or end the session, within a pre-registered horizon. The answer space is discrete; resolution is deterministic from the subsequent observation stream against pre-registered rules; the forecast is timestamped and hashed before the outcome exists.
Scoring follows the standard battery: skill in bits over the R1 population-prior baseline (what wearers in general do at such moments) and over the R2 own-routine baseline (what this wearer's recency-weighted routine predicts), with calibration checked across the forecast set. The permutation specificity gate then swaps in another wearer's egocentric history: if skill does not collapse, the system was reading generic scene priors, not this target.
An A–B design over the same sealed tasks: the assistant with the egocentric feature on versus off (audio-only, or retrieval-only). Same wearers, same forecast horizons, same resolution rules. The difference in skill is the feature's measured contribution — not a demo, not a preference survey.
Egocentric video and audio are the most sensitive data a wearer produces, and none of it needs to leave the company. Raw footage stays inside the operator's boundary; only sealed forecasts, resolved outcomes, and aggregate metrics are exported for scoring.
One leaderboard row per condition: Skill vs R1, Skill vs R2, calibration, and the permutation result — reported per evidence tier, so the marginal lift of L3–L4 vision over L2 audio is a number, not a claim.
Questions the smart-glasses instantiation of TS-Personal is positioned to answer — none of them answerable from capture-quality or recall benchmarks.
Does egocentric vision add calibrated forecasting lift over L2 audio alone — and on which transition types? Vision may pay only where the antecedents are silent: object handling, document context, spatial movement.
Can sparse still images replace continuous video for target-state forecasting? At what sampling rate does skill saturate? The answer sets the power, storage, and privacy budget for the entire category.
Field of view, occlusion, and framing systematically shape what the camera sees of the wearer's context. Can this bias be characterized per device, so that skill comparisons across hardware are meaningful?
Does visibly wearing a camera change the behaviour being forecast — in the wearer and in the people around them? If observation perturbs the target, R2 own-routine baselines drawn from pre-wear periods may misstate skill.
Bits of target-specific skill per joule and per captured byte: on that ledger, does video justify its cost against audio? Evidence-tier ablation turns the trade-off into a measurable quantity.
Wearers can also just say what they intend to do next. Self-report is biased but not dismissible: how much of the egocentric stream's lift survives once a cheap self-report auxiliary channel is included as a baseline?